Research On Urban Road Vehicle Detection And Tracking Based On Deep Learning

Vehicle detection and tracking on urban roads is conducive to improving intelligent traffic management in cities,and is important in relieving traffic pressure in a timely manner and providing a basis for decision making for city managers.However,at this stage,most of the deep learning-based vehicle detection and tracking models are confined to GPUs with large weight files,which are difficult to deploy to mobile terminals.Therefore,in this paper,based on urban road traffic video,do relevant research on vehicle detection algorithms and tracking algorithms from lightweight networks.The main work is as follows:(1)Construct vehicle detection and vehicle tracking datasets.Firstly,the traffic video on the city road is collected,the images are extracted from the video by frame extraction,and Labelimg is used to label the images to build a detection dataset with 3443 images and a total of 24932 vehicle labeled frames;then,the Ve Ri-776 dataset is selected as the vehicle re-identification dataset for the tracking algorithm,which is used to improve the applicability of the algorithm for vehicle tracking;then,the Three videos with a total of 5462 frames from different scenes were selected to create a vehicle tracking evaluation dataset using Darklable software for evaluating the tracking algorithm.(2)Vehicle detection using improved YOLOv5 s based algorithm.Firstly,the lightweight network Shuffle Net V2 was used instead of CSP Darknet53 as the backbone extraction network,and the activation function of Shuffle Net V2 was optimised;then,the GAM attention mechanism with global optimisation function was added to the output side to address the accuracy degradation problem caused by the lightweight network;next,the placement of GAM was ablated The experimental results show that the detection effect is best when the GAM attention mechanism is placed at the second layer of the output layer.The improved algorithm improved the m AP＿0.5 and m AP＿0.5:0.95 metrics by 1.2% and 1.5% respectively over the original algorithm,and the weight file was reduced by 21%.(3)The improved Deep SORT based algorithm was used for vehicle tracking.First,the deep feature extraction network is replaced by a lightweight network;then,to address the problem that the original Deep SORT algorithm is insensitive to vehicle features,the network is retrained using the Ve Ri-776 vehicle dataset to improve the applicability of the algorithm for vehicle target tracking;then,to address the problem that the original algorithm ID Switch is frequent,the IOU metric of detection frame and target prediction frame is proposed to be used Finally,the tracking evaluation metrics are obtained through a Mot challenge file.The results show that the improved Deep SORT algorithm improves the MOTA metric by 2.3%,reduces the weight file by 95% and significantly reduces the number of ID Switches compared to the algorithm before the improvement.(4)Traffic flow statistics based on vehicle detection and vehicle tracking.Firstly,the virtual detection line counting method is used to count the vehicles,by adding multiple detection lines to achieve for multi-directional traffic statistics;then,the vehicle speed detection module is added to achieve the monitoring of vehicle speed;finally,based on Py Qt to develop a video traffic statistics platform,three different environment video is selected to improve the before and after algorithm traffic statistics comparison experiments,the experimental results show that The experimental results show that the improved algorithm is better than the original algorithm in terms of accuracy and speed of traffic flow statistics.In summary,this paper focuses on vehicle detection and tracking in urban road scenarios.The experiments show that the proposed algorithm meets the premise of accuracy,significantly reduces the model weight file,facilitates its deployment to mobile terminal devices,and has certain reference value for intelligent traffic management.